Edge ML Model Adaptation for Hardware-Constrained Devices

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Implementing sophisticated machine learning models on edge devices in aerospace systems is impractical due to hardware constraints, leading to energy-intensive data transmission and long delays in analysis.

Innovation Solution

Adapt machine learning models for target devices using type-specific adaptation methods, including resource mapping and compression techniques, to optimize performance and compatibility.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If machine learning models are implemented on edge devices, then processing speed and energy efficiency are improved, but hardware constraints make sophisticated models impractical to implement

Engineering Contradiction:
Improveprocessing speedVSAvoidhardware constraints
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent transforms machine learning models by changing their structural parameters to suit edge device constraints. This includes converting complex models into simplified versions with reduced computational requirements, adjusted precision levels, and modified architectures that maintain adequate accuracy while fitting within limited hardware resources of edge devices.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent segments sophisticated machine learning models into smaller, more manageable components that can be individually optimized and executed on resource-constrained edge devices. This segmentation allows the system to process complex tasks through multiple simpler steps rather than requiring a single large model, making deployment on edge devices practical.

Inventive Principle:
Principle #1Segmentation

2Reliability

If data is gathered and forwarded to remote processors for analysis, then sophisticated machine learning models can be executed, but energy consumption and communication requirements increase

Engineering Contradiction:
Improvemodel execution capabilityVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts the machine learning processing capability from remote centralized systems and places it directly on edge devices. By taking out the model execution function from the remote processor and embedding it locally on edge devices, the system eliminates the need for continuous data transmission while maintaining sophisticated analysis capabilities at the source.

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If data is transmitted to remote systems for analysis, then comprehensive processing can be performed, but communication delays increase

Engineering Contradiction:
Improveanalysis capabilityVSAvoidcommunication delay
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-processing and analyzing data locally on edge devices using deployed machine learning models before any potential remote transmission. This allows immediate insights and decisions to be generated at the edge, with only necessary summarized information or anomalies being transmitted remotely, thereby eliminating most communication delays.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4672086A1Machine learning models
Publication Date: 2025.12.31 ARINC INC
  • EP4672086A1 patent drawingFigure 1~2
  • EP4672086A1 patent drawingFigure 3~4
  • EP4672086A1 patent drawingFigure 5~6

AI summary

A computer implemented method of adapting a machine learning model for execution on a target device (102) is provided. The method comprises receiving information describing a trained machine learning model (501, 605, 701); determining a type of the machine learning model; receiving an indication (503, 607) of one or more characteristics of the target device; adapting the machine learning model for the target device using an adaptation method (306, 308) that is selected based on the type of the machine learning model; and outputting information (516, 622, 712) describing the adapted machine learning model.